Wind Energy, the Price of Carbon, and Carbon Emissions: Evidence from Ireland Kevin F. Forbes Department of Business and Economics The Center for the Study of Energy and Environmental Stewardship The Catholic University of America Washington, DC USA Forbes@CUA.edu Marco Stampini1 Inter-American Development Bank Washington, DC USA Ernest M. Zampelli Department of Business and Economics The Center for the Study of Energy and Environmental Stewardship The Catholic University of America Washington, DC USA 30th USAEE/IAEE North American Conference Oct 9-11 2011 Washington DC 1 This paper reflects the opinions of the authors and not those of the Inter-American Development Bank, its Board of Directors, or the countries they represent. Abstract This paper examines the effect of wind energy and outcomes in the European carbon market on the carbon levels emitted by the power grid in the Republic of Ireland over the period 27 March 2008 through 15 August 2010. Using half-hour data on actual wind energy production, forecasted wind energy production, forecasted load, actual load, the price of natural gas relative to price of coal, the price of carbon allowances, the ex ante system price of electricity, and the actual levels of carbon emissions, a model is estimated that examines the contribution of wind energy, both load and wind forecasting errors, fuel prices, and the carbon price on carbon emissions. The results of the analysis are expected to be of interest to anyone with an interest in energy and climate policy. Keywords: Climate Change, Renewable Energy, Wind Energy, Wind Forecasting JEL Codes: Q2, Q4, Q5 1. Introduction Because of the environmental concerns associated with fossil fuels use, considerable support exists for increasing the share of electricity generation attributable to wind. For example, on April 12, 2011, CaliforniaGovernor Edmund G. Brown, Jr. signed legislative bill SBX1 2 which requires that one-third of California’s electricity come from renewable sources by 2020. This policy initiative is expected to increase the wind energy capacity available to the California Independent System Operator (ISO) by almost 500 per cent by 2020 (Hawkins, 2008). The European Union has established a goal of 20 percent renewable energy by 2020 with wind energy serving as a key source of the increase. And in China, whose installed wind capacity is fifth largest in the world, wind power is expected to contribute 10 percent to its total electricity needs by 2020 (Chen, et al., 2009). These proposed increases in wind energy are predicated implicitly on the belief that an increase in wind energy is an effective mechanism to reduce carbon emissions. But is it? According to wind energy advocates, the answer is a definite “yes” since thereare no direct carbon emissions associated with the production of electricity from wind turbines while the production of electricity using coal gives rise to approximately one ton of CO2 emissions per MWh. Consistent with this view, Cullen (2008) has presented evidence that eachMWh of wind power in ERCOT, the power grid serving the vast proportion of Texas, offsets approximately one ton of CO2emissions. However, the finding is open to question in light of evidence that wind energy largely displaces natural gas, a fuel with a carbon emissions factor that is substantially less than one ton per MWh. Moreover, the study fails to address a potentially important point-the relative unpredictability of wind energy may actually contribute to an increase in carbon emissions at conventional power plants. For example, to maintain the reliability of the power system, an over-prediction of wind energy levels may induce the system operator to dispatch a generating unit that is highly carbon intensive while an under-prediction of wind energy may increase the carbon intensity of the conventional plants that are dispatched down to accommodate the unexpected wind energy production. While most environmental economists tend to favor policy initiatives that would place a price on carbon, the California energy crisis of 2000/2001,the oil price spike of 2007/2008, and the 2008/2009 financial crisis, have contributed to a reluctance to rely on market processes in remedying energy market inefficiencies. Increasingly, many are of the view that while markets may “work” in the abstract, they are frequently deficient in practice. For example, the European carbon trading system has been sharply criticized. Indeed, a 2009 study conducted by the Friends of the Earth organization in the United Kingdom has concluded that Europe’s carbon trading system is ineffective in reducing carbon emissions. 2. The Irish Power Grid This paper examines the effect of wind energy and outcomes in the European carbon market on the carbon levels emitted by the power grid in the Republic of Ireland over the period 27 March 2008 through 15 August 2010. Wind energy accounted for approximately 9.3 percent of load over this period. The data for the analysis were obtained largely from EirGrid (http://www.eirgrid.com), the system operator the Republic of Ireland. Electricity data was also obtained from SEMO (Single Electricity Market Operator), which is a joint venture of EirGridand the SONI (System Operator for Northern Ireland). As of November 2007, the overall coordination of electricity production and consumption on the island is conducted by SEMO (http://www.sem-o.com/Pages/default.aspx ). Table 1 reports the 2009 electricity supply capacity for the entire island by fuel type. Table 1 2009 Electricity Supply Capacity in Ireland by Fuel Type in Gigawatts (GW) Wind Other Renewables Coal/Oil CCGT OCGT Interconnector Pumped Storage 1.5 0.6 3.2 3.5 0.7 0.5 0.3 Source: Pöyry Energy EirGrid is one of the few system operators that reports carbon emissions in real-time on its web site. The reported emissions are based on the fuel mix and the heat rates of the generating units. Figure 1 reports the emissions for the first full week of 2010. Carbon intensity was approximately 538 grams/KWh over the period27 March 2008 through 15 August 2010. To put this number in perspective, the carbon intensity of the overall power grid in the United States was approximately 600grams/KWh in 2008, the last year for which data are available. The Irish government has a goal of reducing carbon intensity to 100 grams per KWh. Increasing the share of generation from renewables such as wind energy is viewed as an important component of this strategy. Specifically, wind energy capacity is expected to increase to at least 7.9 GW by 2035 (Pöyry Energy, 2010, p 52). Figure 1. Carbon Emissions from the Republic of Ireland’s Power Grid, January 3 – 9, 2010 2000 1800 1600 Metric Tons Per Hour 1400 1200 1000 800 600 400 200 0 Source: EirGrid EirGrid and SEMO also report actual wind energy production, forecasted wind energy production, actual load, and forecasted load for the Republic of Ireland. Using this data, the forecast accuracy of wind energy can be compared with that of load. The RMSE of the wind forecastsis 29.4 percent of the mean level of wind energy production which is over seven times the RMSE of the load forecasts relative to the mean level of load (Figure 2). The large differential in the RMSEs is consistent with the findings of Forbes, et al. (2011) who have reported large differences in the RMSEs between wind and load for a number of control areas including Western Denmark, Eastern Denmark, 50Hertz in Germany, the Amprion system in Germany,Terna in Italy, National Grid in the United Kingdom, the Midwest ISO in the USA, and ERCOT in the USA. It may be noted that the RMSE of 29.4 percent appears much larger than the six percentwind forecasting error for Ireland that was reported by Lang (2006). But the calculation by Lang is weighted by wind energy capacity which we believe to be inappropriate. Unfortunately, a number of other researchers including Cali, et. al (2006), Krauss et. al. (2006), Holttinenet. al. (2006),Kariniotakiset. al, (2006),and Porter and Rogers (2010, p. 5)have also chosen to report only capacity weighted RMSEs. Figure 2. The Root-Mean-Squared-Errors of the Wind and Load Forecasts in Republic of Ireland’s, 27 March 2008- 15 August 2010 35% RMSE as a Percent of the Mean Level of Wind and Load, Respectively 30% 25% 20% 15% 10% 5% 0% Wind Load Based on an analysis of 41,577 half-hour observations. The data were obtained from EirGrid and SEMO. The forecast errors are calculated based on the one-hour ahead wind forecasts and four day ahead load forecasts. 3. A Model of Carbon Emissions The amount of CO2 emitted by a power grid is determined not just by actual system demand (load) and actual wind energy production but also by how closely these match up with what the system operator had expected demand and wind energy production to be, as unexpected changes in load and/or wind energy may force the system operator to undertake actions that increase (or decrease) CO2 emissions from its conventional generating units. Additionally, changes in relative fuel prices (natural gas and coal) and changes in the price of carbon (if applicable) will have an impact on the optimal mix of generating units chosen by the system operator and hence on the level of CO2 emissions as well. Consequently, we model carbon emissions as a function of forecasted load, positive and negative load forecast errors, forecasted wind energy production,positive and negative wind energy forecast errors, the spot price of natural gas relative to the market price of coal, and the price of carbon relative to the ex ante system electricity price established by SEMO. We include binary variables to control for each hour of the day and for each month of the year. In its most general algebraic form, the model is given by: ln CO2 f ( forecasted load , one hour ahead wind forecast , positive wind forecast error , negative wind forecast error , positive load forecast error , negative load forecast error , gas / coal price rati o, carbon/electric p rice ratio , hourh , monthm ) where: lnCO2 is the natural logarithm of CO2 emissions measured in metric tons per hour; forecasted load is the forecasted level of electricity demand in MW; one hour ahead wind forecastis the one hour ahead forecasted level of wind energy; (1) positive wind forecast errorequals the absolute value of the difference between the forecasted and actual levels of wind energy when the forecasted level of wind energy is greater than the actual and zero otherwise; negative wind forecast error equals the absolute value of the difference between the forecasted and actual level of wind energy in period t when the forecasted level of wind energy is less than the actual and zero otherwise; positive load forecast errorequals the absolute value of the difference between the forecasted and actual level of demand when the forecasted level of demand is greater than actual and zero otherwise; negative load forecast errorequals the absolute value of the difference between the forecasted and actual level of demand when the forecasted level of demand is less than actual. It is zero otherwise; gas/coal price ratioequals the spot price of natural gas relative to the price of coal where both prices are expressed in EUR per MMbtu; carbon/electric price ratio equals the European carbon price relative to the half-hour ex ante system price of electricity established by SEMO. houris a vector of binary variables representing each hour of the day exclusive of hour one. Hour 1 is reflected in the overall constant term; monthis a vector of binary variables representing each month of the year exclusive of month 1, January. January is reflected in the overall constant term. Data for each variable is reported for every thirty-minute period from 27 March 2008 to 15 August 2010. The initial estimation of (1) was conducted using the multivariable fractional polynomial (MFP) model, a useful technique when one suspects that some or all of the relationships between the dependent variable and the explanatory variables are non-linear (Royston and Altman, 2008), but there is little or no basis, theoretical or otherwise, on which to select particular functional forms. The MFP approach begins by estimating a model that is strictly linear in the explanatory variables. In subsequent iterations, the algorithm cycles through a battery of nonlinear transformations of the explanatory variables (positive and negative powers, natural logarithms, etc.) until itconverges to the combination of functional forms that best predicts the dependent variable. In our case, the analysis provided support for including nonlinear forms for three of the explanatory variables, exclusive of hour and month binary variables. Specifically, the MFP model obtained is given by: ln CO2 0 1 ln( forecasted load ) 2 (one hour ahead wind forecast ) 3 ( positive wind forecast error ) 4 (negative wind forecast error ) 5 ( positive load forecast error ) 6 (negative load forecast error ) (2) 7 ( gas / coal price rati o) 3.0 8(carbon/ele ctric pric e ratio ) 0.5 24 12 h2 m2 h hourh m monthm The hypothesized signs for each of the model’s coefficientsand their rationales are presented in Table 2. Table 2 Hypothesized Signs for Model Parameters Variable Coefficient Hypothesis ln(forecasted load) β1 >0 one hour ahead wind forecast β2 <0 positive wind forecast error β3 >0 negative wind forecast error β4 <0 positive load forecast error β5 <0 negative load forecast error β6 >0 (gas/coal price ratio)3.0 β7 >0 (carbon price ratio)-0.5 β8 <0 Rationale More electricity consumption increases electricity production from fossil fuels More wind energy reduces electricity production from fossil fuels Lower than expected wind energy increases reliance on electricity produced using fossil fuels Higher than expected wind energy reduces reliance on electricity produced using fossil fuels Lower than expected electricity demand reduces electricity production using fossil fuels Higher than expected electricity demand increases electricity production using fossil fuels Higher relative price for natural gas makes electricity production from natural gas less economic Higher relative carbon price makes electricity production from fuels with higher carbon emissions (e.g.coal) less economic Equation 2 was re-estimated using least squares with standard errors robust to arbitrary forms of heteroscedasticity and autocorrelation.2 The results are reported in Table 3. For convenience, we do not report the estimated coefficients for the hour and month binary variables. Table 3 Parameter Estimates for MFP Functional Form with Robust Standard Errors Variable Est. Coef Std. Error ln(forecasted load) one hour ahead wind forecast positive wind forecast error negative wind forecast error positive load forecast error negative load forecast error 1.479533 -0.00049 0.000497 -0.00053 -0.00059 0.000533 (gas/coal price ratio)3.0 0..5 (carbon price ratio) constant Z-Value P-Value 0.021507 1.08E-05 3.84E-05 3.92E-05 2.21E-05 2.76E-05 68.79 -45.08 12.95 -13.63 -26.66 19.31 0.000 0.000 0.000 0.000 0.000 0.000 0.012392 0.000478 25.95 0.000 0.015422 -5.17502 0.005467 0.169579 2.82 -30.52 0.005 0.000 Inspection of Table 3 reveals that all parameter estimates are highly statistically significant, with p-values well below 0.01, and all are consistent with the hypothesized signs in Table 2. Reductions in CO2 emissions can be expected with reductions in forecasted load and increases in expected wind power. To assess quantitatively the impact of a one MWh reduction in load, first note that: ln CO2 CO2 forecasted load ln( forecasted load ) ( forecasted load ) CO2 This means that: 2 Augmented Dickey-Fuller tests were conducted for all variables. The null hypothesis of a unit root was rejected in all cases. Results are available upon request. CO2 ln CO2 CO2 ( forecasted load ) ln( forecasted load ) forecasted load (3) Evaluating (3) at the sample mean values for CO2 and forecasted load, and using the (rounded) point estimate of β1from Table 3 yields: CO2 828.5 1.48 0.402 ( forecasted load ) 3050.3 This in turn must be divided by two since forecasted load in the sample is measured in MW per half-hour period. Consequently, a reduction in forecasted load of one MWh can be expected to reduce emissions by approximately 0.2 metric tons per hour. This, of course, presumesno change in load forecasting errors. Specifically, eachMWh increase in positive (negative) load forecasting errors will yield subsequent reductions (increases) in CO2 emissions of 0.24 (0.22) metric tons per hour.3 In calculating the expected emissions impact of a one MWh increase in forecasted wind power, we note first that: CO2 ln CO2 CO2 (one hour ahead wind forecast ) (one hour ahead wind forecast ) , This implies that a one MWh increase in forecasted wind energy production can be expected to reduce CO2 emissions byabout 0.20metric tons per hour, presuming no change in wind forecasting errors.4 A MWh increase in positive (negative) wind forecasting errors will generate 3 The calculation proceeds as follows: CO2 ln CO2 CO2 ( positive load forecasting error ) ( positive load forecasting error ) 0.00059 828.5 0.48 evaluated at the sample mean value of CO2. Dividing by two yields the -0.24 reported above. Again, division by two is required since the forecasted wind data is in MW per half-hour period. 4 This is equal to (-0.00049×828.5)/2. an accompanying increase (decrease) of about 0.21 (0.22) metric tons of CO2 per hour. The results in Table 3 are consistent with our hypothesis that an increase (decrease) in the price of natural gas relative to coal will serve to increase (decrease) carbon emissions, ceteris paribus. Additionally, they are supportive of the hypothesis that an increase in the carbon price relative to the electricity price will reduce emissions of CO2. Quantitatively, our estimates suggest that a one percent decline in the relative price of natural gas reduces carbon emissions by about 1.4 metric tons per hour while a one percent increase in the relative carbon price reduces emissions by 0.11 metric tons per hour.5 3.1 ARCH Effects Time series regression with high frequency data—hourly, daily, weekly—can often be plagued by the problem of time-varying volatility, more commonly referred to as Autoregressive Conditional Heteroscedasticity (ARCH). In such instances, the variance (or volatility) of the random disturbance term in time period t depends on the magnitudes of the random disturbances from past periods. Algebraically, this can be represented by: ht 0 1et21 2 et22 ... q et2q (1) or, in shorthand, an ARCH(q) model, where ht is the error variance in time t and et2 s is the squared error in time t-s. To test for ARCH effects, the basic model is estimated, the residuals retrieved, and the following equation is estimated using least squares: eˆt2 0 1eˆt21 2 eˆt22 ... q eˆt2q where êt is the residual for time period t. Significance of any αj indicates the presence of ARCH effects. 5 Please refer to Appendix A for the derivations upon which these calculations are based. Since our model does employ high-frequency time-series data, we tested for ARCH effects using the residuals from the MFP estimation including up to 24 lags. A number of the lagged squared residuals were statistically significant, though the marginal impact of the first period lag clearly dominated with a point estimate of approximately 0.90, twenty to thirty times the magnitudes of the other estimated coefficients. Given the presence of significant ARCH effects, the model was re-estimated, with the same MFP functional form, using the Generalized ARCH (GARCH) procedure. GARCH is generally used to account for long lagged effects with a relatively small number of parameters. Specifically, rewrite (1) as: ht 0 1et21 11et22 121et23 ... 1q11et2q effectively imposing a geometric lag structure where s 1s 1 1et2 s . Now add and subtract β1α0 and rearrange terms on the right hand side to yield: ht ( 0 1 0 ) 1et21 1 ( 0 1et22 121et23 ... 1q11et2q ) (2) The last term in parentheses is nothing more than ht-1 and hence (2) can be written as: ht 0 1et21 1ht 1 This is the standard GARCH(1, 1) model with one lagged h term and one lagged squared error term. Its popularity derives from the fact that it can capture similar effects as an ARCH(q) model with only three parameters rather than (q + 1) parameters. The GARCH(1, 1) estimates are shown in Table 4. A comparison with Table 3 reveals virtually the same results except for more precise standard errors and higher Z-values as a consequence of modeling directly the ARCH process. Point estimates are of similar magnitudes and hence lead to the same quantitative assessments as already discussed in the previous subsection. We note also that both the lagged squared error and the lagged error variance are statistically significant in the estimated error variance function. Table 4 GARCH(1, 1) Estimates--MFP Functional Form Variable Est. Coef Std. Error Z-Value P-Value ln(forecasted load) one hour ahead wind forecast positive wind forecast error negative wind forecast error positive load forecast error negative load forecast error 1.414419 -0.00047 0.000476 -0.00056 -0.00056 0.000496 0.001709 1.62E-06 7.15E-06 6.77E-06 4.47E-06 6.28E-05 775.72 -293.25 65.15 -87.48 -126.04 120.45 0.000 0.000 0.000 0.000 0.000 0.000 (gas/coal price ratio)3.0 0.012141 7.14E-05 117.8 0.000 (carbon price ratio-0.5 Constant GARCH(1, 1) lagged squared error lagged variance Constant 0.015668 -4.65938 0.000922 0.023626 25.38 -197.22 0.000 0.000 0.894295 0.044384 0.001006 0.017338 0.00562 1.94E-05 51.58 7.9 52 0.000 0.000 0.000 Finally, as a robustness check, we estimated two other versions of the model employing GARCH(1, 1). The first assumed a log-linear form (dependent variable in natural log with right hand side variables in linear form) while the second assumed a double-log form (all variables in natural log form, except for the “error” variables). The results, reported in Appendix B, are consistent with those reported in the text. The one exception is that the marginal impact of an increase in the relative price of carbon seems to be greater for the alternative functional forms. 4. Summary and Conclusions This paper has examined the effect of wind energy on the carbon levels emitted by the power grid in the Republic of Ireland over the period 27 March 2008 through 15 August 2010. Using half-hour data on forecasted and actual wind energy production, forecasted and actual load, the price of natural gas relative to price of coal, the price of carbon allowances relative to the ex ante system price of electricity, and the actual levels of carbon emissions, we estimated a model that examined the contribution of wind energy, both load and wind forecasting errors, fuel prices, and the carbon price on carbon emissions. Based on the sample averages, the results suggest that a one MWh decrease in load and a one MWh increase in wind power will reduce carbon emissions by almost the same amount, as long as there are no offsetting changes in load and or wind forecasting errors. It is instructive to note that our estimate of this reduction is far lower than that reported by Cullen (2008). As noted in the introduction, Cullen’s analysis suggests that a one MWh increase in wind power reduces emissions by one metric ton. The estimates presented in this paper suggest something much smaller, on the order of only 0.2 metric tons. Moreover, based on the MFP suggested functional form (where forecasted load is represented in the estimating equation in terms of its natural logarithm while forecasted wind is modeled in terms of its actual level), there are diminishing marginal returns to increases in wind energy production on carbon emissions. Accordingly, increases in wind energy penetration above the current level may have only a minor impact on the level of carbon emissions. In contrast, our results seem to suggest that fairly small decreases in the price of natural gas relative to coal and fairly small increases in the relative price of carbon can be very effective in reducing emissions of CO2. This of course implies that the imposition of carbon taxes or the establishment of well designed carbon cap-and-trade programs might be more cost-effective ways of reducing carbon emissions than systems of wind power subsidies and tax credits. Appendix A To determine the change in CO2 emissions from a one percent change in the gas/coal price ratio, we derive: CO2 ln CO2 CO2 ( gas / coal price ratio) ln( gas / coal price ratio ) ( gas / coal price ratio) 3 ˆ ( gas / coal price ratio) 2 CO ( gas / coal price ratio) 7 2 3 0.012392 1.667 828.5 1.667 142.34 2 using sample mean values. The value for this derivative suggests that a one percent increase (decrease) in the gas/coal price ratio will cause a (142.34/100) or 1.4234 metric ton increase (decrease) in CO2 emissions. Similarly for the change in CO2 emissions from a one percent change in the carbon/electricity price ratio, we derive: CO2 ln CO2 CO2 (carbon price ratio) ln( carbon price ratio) (carbon price ratio) 0.5 ˆ8 (carbon price ratio) 1.5 CO2 (carbon price ratio) 0.5 0.015422 0.332 1.5 828.5 0.332 11.0822 Correspondingly, the value suggests that a one percent increase in the carbon/electricity price ratio will cause a 0.11 metric ton decrease (increase) in CO2 emissions. Appendix B Table B1 GARCH(1, 1) Estimates--Log-Linear Functional Form Variable Est. Coef Std. Error Z-Value P-Value forecasted load one hour ahead wind forecast positive wind forecast error negative wind forecast error positive load forecast error negative load forecast error 0.000442 -0.00044 0.000467 -0.00053 -0.00057 0.000501 1.09E-06 1.54E-06 6.52E-06 6.36E-06 3.20E-06 4.12E-06 404.61 -284.42 71.64 -83.01 -177.56 121.5 0.000 0.000 0.000 0.000 0.000 0.000 gas/coal price ratio 0.129848 0.000921 140.99 0.000 carbon price ratio constant GARCH(1, 1) lagged squared error lagged variance constant -0.12882 5.224772 0.003667 0.004884 -35.13 1069.81 0.000 0.000 0.895158 0.048036 0.048036 0.017438 0.005784 0.005784 51.33 8.31 8.31 0.000 0.000 0.000 Table B2 GARCH(1, 1) Estimates--Double Log Functional Form Variable Est. Coef Std. Error Z-Value P-Value ln(forecasted load) ln(one hour ahead wind forecast) positive wind forecast error negative wind forecast error positive load forecast error negative load forecast error 1.39459 -0.0966 0.0005 -0.0006 -0.0005 0.00048 0.00336 0.00041 6.71E-06 6.68E-06 4.26E-06 4.01E-06 415.12 -234.99 74.55 -84.85 -126.23 119.48 0.000 0.000 0.000 0.000 0.000 0.000 ln(gas/coal price ratio) 0.22154 0.00158 140.04 0.000 ln(carbon price ratio) constant GARCH(1, 1) lagged squared error lagged variance constant -0.0351 -4.1514 0.00107 0.0266 -32.73 -156.04 0.000 0.000 0.91009 0.04152 0.001 0.01757 0.00565 2E-05 51.8 7.35 48.98 0.000 0.000 0.000 References Bentek Energy LLC (2010). How less became more: Wind, power and unintended consequences in the Colorado energy market. Technical report, Bentek Energy LLC. Cali, Ü., B. Lange, R. Jursa, and K. Biermann, (2006), Short-term prediction of distributed generation – Recent advances and future Challenges, ElftesKasseler Symposium EnergieSystemtechnik November 2006. Available on the Internet at http://www.iset.unikassel.de/public/kss2006/KSES_2006.pdf Cullen, Joesph (2008) What's Powering Wind? Measuring the Environmental Benefits of Wind Generated Electricity, paper prepared for presentation at the American Agricultural Economics Association Annual Meeting, Orlando, FL, July 27-29, 2008. Forbes, K., M. Stampini, and E. Zampelli (2011) Are Policies to Encourage Wind Energy Predicated on a Misleading Statistic? unpublished manuscript. GregorGiebel, PoulSørensen, HanneleHolttinen, Forecast error of aggregated wind power (2007) Giebel, G., Landberg, L., Kariniotakis, G., Brownsword, R.,"State-of-the-Art on Methods and Software Tools for Short-Term Prediction of Wind Energy Production", CD-Rom Proceedings of the European Wind Energy Conference & Exhibition EWEC 2003, Madrid, Spain, June 16-19, 2003 Giebel, G., Kariniotakis, G., Brownsword, R., "The State-of-the-Art in Short-Term Prediction of Wind Power - From a Danish Perspective", CD-Rom Proceedings of the 4th International Workshop on Large-Scale Integration of Wind Power and Transmission Networks for Offshore Wind farms, Billund, Denmark, October 20-21, 2003. Holttinen, H., P. Saarikivi, S. Repo, J. Ikäheimo, and G. Koreneff: Prediction Errors and Balancing Costs for Wind Power Production in Finland. Global Wind Power Conference, Adelaide, 2006 Krauss, C, B. Graeber, M. Lange, U. Focken: Integration of 18GW Wind Energy into the Energy Market – Practical Experiences in Germany. Workshop on the Best Practice in Short term Forecasting, Delft (NL), 25 October 2006. Kariniotakis G., "State of the art in wind power forecasting", 2nd International Conference on Integration of Renewable Energies and Distributed Energy Resources, Napa, California/USA, 48 December 2006. Katzenstein, W. and J. Apt (2009).Air emissions due to wind and solar power.Environmental Science and Technology 43 (2), 253{258. Lange, B., K. Rohrig, B. Ernst, F. Schlögl, Ü. Cali, R. Jursa, and J. Moradi: Wind power prediction in Germany – Recent advances and future challenges. European Wind Energy Conference and Exhibition, Athens (GR), 27.2.2.3. 2006 Lange, B., Wessel A., Dobschinski J., Rohrig,K. Role of Wind Power Forecasts in Grid Integration KASSELER SYMPOSIUM ENERGIE-SYSTEMTECHNIK 2009, p 118-130 http://www.iset.uni-kassel.de/public/kss2009/2009_KSES_Tagungsband.pdf Milligan, Michael, Porter Kevin, DeMeo Edgar, Denholm Paul HolttinenHannele, Kirby Brendan, Miller Nicholas Mills Andrew O’Malley Mark, Schuerger Matthew, and SoderLennart (2009) Wind Power Myths Debunked, IEEE Power and Energy, November/December vol 7 no 6, pp. 89-99. NERC, (2009), Accommodating High Levels of Variable Generation, http://www.nerc.com/files/IVGTF_Report_041609.pdf NERC (2010) Variable Generation Power Forecasting for Operations http://www.nerc.com/files/Varialbe%20Generationn%20Power%20Forecasting%20for%20Oper ations.pdf Nielsen, T.S., H. Madsen, H. Aa. Nielsen, P. Pinson, G. Kariniotakis, N.Siebert, I. Marti, M. Lange, U. Focken, L. von Bremen, P. Louka, G. Kallos, G. Galanis: Short term Wind Power Forecasting Using Advanced Statistical Methods. European Wind Energy Conference and Exhibition, Athens (GR), 27.2.2.3. 2006 Pöyry Energy, (2010), Low Carbon Generation Options for the All-Island Market . Technical report to Eirgrid.http://www.eirgrid.com/aboutus/publications/